Abstract: Bridge cable damage detection is a critical aspect of bridge safety operation and maintenance. How to quickly and automatically process cable images and accurately detect damaged areas is key to ensuring bridge safety operations.This paper improved the large-scale segmentation model SAM (segment anything model)and applied it to the semantic segmentation of bridge cable damage,providing significant evidence for damage detection.The improvements to SAM consisted of two main points:①Fine-tuning an Adapter on the image encoder and enhancing the model's generalization to bridge cable data through transfer learning methods.②Fine-tuning the prediction head of the mask decoder to enable SAM to perform semantic segmentation without prompts and introduce multi-class capabilities.To verify the advantages of the improved SAM,comparative experiments had been conducted using a dataset of nearly 2200 bridge cable images against DeepLabV3+.The results showed that the improved SAM performed better in scenarios with imbalanced damage category distributions and limited sample sizes.Its semantic segmentation evaluation metrics included a mean intersection over union (mIoU)of 73.41%,an average F1 score of 83.99%,and mean class accuracy of 81.80%.

Key words: deep learning, semantic segmentation, large model fine-tuning, damage detection of bridge cable